#33 · Systems biology

Thermodynamic penalties, learned posteriors and adaptive ABC for mechanistic inference

The common thread is methodological: each item modifies how Bayesian or likelihood-based inference is performed for a mechanistic biological model, with either physical constraints or computational cost driving the design. For mass-action biochemical reaction networks, one preprint penalizes detailed-balance violations during both parameter optimization and MCMC-based Bayesian inference. For phylogenetics, a second preprint trains conditional generative models over trees with discrete diffusion and reports more accurate recovery of MCMC posteriors than Phyloformer 2 on the datasets considered. A third applies adaptive minibatch SMC ABC, using small batches early and growing them as posteriors refine, to large heterogeneous biological datasets. What is new is soft thermodynamic penalties in network inference, learned conditional posterior approximators for phylogenetics, and batch-size adaptation within ABC. The application domains differ, so the shared thread is methodological rather than a common dataset or model class.